CCTV Weapon Detection: Rifles vs Umbrellas

False positives are the single biggest obstacle to deploying weapon detection AI in the real world. This dataset directly addresses the problem through Hard Negative Mining: it contains a balanced 50/50 split of images with people carrying rifles (real threats) and people carrying umbrellas (common false-positive triggers). Both object types share a similar elongated silhouette from CCTV angles, forcing models to learn the subtle visual differences. The dataset includes varied lighting (daylight, dusk, night), weather conditions, sensor noise, and motion blur. Every image is annotated in YOLO format with classes for person, rifle, and umbrella. The open-source sample provides 120 images; the full package includes 1,000 images with 8 evaluation videos.